| Item |
Value |
| Date |
2026-09-25 |
| Environment |
Windows 11 x64; rustc 1.98.0; Node v24.21.0 / npm 12.0.2 |
| Code revision |
Uncommitted tree: accepted M2-B/C plus this work |
| Design references |
Adapter layout, unknown versions, M2 migration, validation.md V17/V30 |
| Execution |
Main session implemented Codex/framework/storage as the reference; five parallel subagents migrated Pi/OMP/Claude/Gemini/Qwen; main session integrated and validated |
Commands and results
| # |
Command/directory |
Exit |
Result |
| 1 |
cargo test –locked -p llm-usage-core, desktop/src-tauri |
0 |
276 passed/0 failed: M2-B/C baseline 249 plus net 27, covering V17 fallback, version samples, directories, registries |
| 2 |
cargo test –locked –workspace, desktop/src-tauri |
0 |
278 passed: core 276, app 2 |
| 3 |
cargo clippy –locked -p llm-usage-core –all-targets – -D warnings |
0 |
Passed; existing app main.rs test-module ordering warning outside this core check, inherited from M0 |
| 4 |
npm run verify, repository root |
0 |
Markdown 94 files/0 issues after fixing existing SKILL.md MD060 table padding, also failing at HEAD; assets/scripts/Svelte/fmt/Clippy/Rust/build pass; Vite 518.01 kB, gzip 176.33 kB |
| 5 |
cargo run -p llm-usage-core –example real_verify_codex against ~/.codex, output build/…/real-check-codex-m2d |
0 |
Native results below |
| 6 |
real_verify examples for Pi/OMP/Claude/Gemini/Qwen |
0 |
Pi/OMP exactly match m2bc-resumed values; Claude/Gemini/Qwen remain no_data/not_found |
Directory migration: V30 structure
- Six agents moved into desktop/src-tauri/crates/core/src/adapters/agent/: mod.rs stable
entry and SourceAdapter implementation; detect.rs detection/version selection;
versions/mod.rs registry with VERIFIED_VERSION_IMPLS, LATEST_IMPL_ID and shared select();
versions/format_id.rs format implementation. Removed root codex.rs/claude.rs/pi.rs/
omp.rs/gemini.rs/qwen.rs.
- Product mappings moved to codex/common.rs and claude/common.rs. usage_map.rs retains
cross-agent map_pi_family for Pi/OMP and map_genai_usage for Gemini/Qwen, plus future
Kimi/ZCode/Copilot/Kilo mappings awaiting M3–M5 adapters.
- tests/adapter_layout_v30.rs checks directories/entries/registries, absence of root
single-file implementations and product mappings, and preservation of shared mappings.
- Public entry/source identity/cursors remain compatible. Format parse contexts add
version_basis with serde(default); Codex/Claude/Pi/OMP each test old-context deserialization
without cursor reset. file_identity/generation/cursor formats unchanged.
Unknown-version reads: New V17 behavior
- DetectOutcome::Supported adds basis: VersionBasis, known_version/latest_fallback,
and nullable format_version. UnsupportedVersion remains only for confirmed incompatible
versions: fixed source establishes a different saved layout for Pi v1/v2/missing version,
so these do not attempt fallback.
- Schema v3 persists compatibility information; parse_basis is excluded from content hashes:
usage_events.parse_basis per event; source_files.format_status JSON with format/
found_version/basis/compat; latest_fallback diagnostic once per file/generation;
active_compat file status.
- V30 fallback failure: records read, zero events and structural diagnostics means incompatible.
No event/cursor/aggregate commit; next run may retry. Partially usable files retain validated
records and diagnostics describing incomplete coverage.
- Tests for schema<3 freeze the old column set without persisting parse_basis, matching
historical behavior. Normal database opening always migrates first.
- Revised tests replace blanket unknown-version rejection with marked successful fallback
for Codex/Pi/OMP; identifiable Codex without version; added optional fields; damaged structure
rejected while retaining old results; partial Codex reads; confirmed incompatible Pi rejected.
Per-version Codex samples: Remaining M2-A work
Ignored extraction tool build/desktop-usage-validation/tools/extract-codex-versions.mjs
chooses the smallest usage-containing file per version. Redaction preserves numbers,
replaces strings with REDACTED, IDs with anon-N and cwd with <PATH>. jq independently
calculates expected values. Leak checks find no UUIDs/private paths/message text.
| Version |
Representative file |
Calls |
Total sum |
Snapshot comparison |
Result |
| 0.153.0 |
276 lines |
25 |
934,367 |
matched |
rollout_v1 verified; sample added to the repository |
| 0.154.0-alpha.6.1 |
37 lines |
3 |
69,740 |
matched |
rollout_v1 verified; sample added to the repository |
| 0.154.0-alpha.6.2 |
36 lines |
5 |
115,757 |
matched |
rollout_v1 verified; sample added to the repository |
| 0.155.0-alpha.16.3 |
M2-A three sessions |
49 and others |
— |
— |
Verified in M2-A |
Native finding: all inspected local 0.139–0.151 files lack token_usage_record. They
contain event_msg/token_count cumulative snapshots: for example, a 0.142.5 session has
74 token_count records and no individual usage records. Unlike rollout_v1, these formats
remain LatestFallback and then incompatible, with visible diagnostics, no invented data
and no cursor advancement. A dedicated legacy parser requires verification of stable
last_token_usage record identity and call boundaries first; these versions were not registered
as verified. codex_versions_v17.rs tests named legacy_carrier_versions_stay_fallback_* fix this behavior.
Native read-only Codex check, 2026-09-25
| Metric |
After M2-D |
M2-A baseline, 2026-09-24 |
| Complete files |
61: 53 verified 0.155/0.154/0.153 files plus 8 newly registered version files |
53, only 0.155.0-alpha.16.3 |
| Incompatible legacy 0.139–0.151 files |
238; previously 212 unsupported_version rejections |
212 |
| Events |
2,578 |
2,620 |
| Repeat scan |
rescan_added=0, no duplicates |
Same |
| UTC totals |
calls 2,578; input_total 295,984,414; cache_read 275,255,936; output 1,502,259; total 297,486,673 |
calls 2,620; total 269,817,926 |
Lower event counts reflect source retention changes. Only four 0.155.0-alpha.16.3 files
remained locally; M2-A’s 53 complete scans included more files from that version before
Codex cleaned old rollouts. Newly registered versions add some events. Regression tests
and exact Pi/OMP before/after results separately verify that migration preserves parsing.
Outstanding work at this stage
| Item |
Status |
Next step |
| Dedicated Codex 0.139–0.151 token_count/last_token_usage parser |
Not implemented |
Verify stable individual-record identity/call boundaries, then add codex/versions implementation and per-version samples |
| Native Claude/Gemini/Qwen samples |
No local data |
Obtain/redact/recheck when available |
| Automatic fallback-file replay after parser upgrade |
Not implemented; adapter limitations record this |
Explicit rescan, or later trigger on parser_version changes |
| M6 source-page compatibility status |
Not implemented |
M6 |
Validation files
- adapters/{codex,claude,pi,omp,gemini,qwen}/, adapters/framework.rs,
storage/schema.rs v3 and ingest.rs.
- tests/adapter_layout_v30.rs, tests/codex_versions_v17.rs and adapter *_gaps_synthetic.rs
updated for V17 behavior.
- tests/fixtures/codex/rollout-v{0.153.0,0.154.0-alpha.6.1,0.154.0-alpha.6.2}.*.
- Ignored build/desktop-usage-validation/: codex-versions extracted files/manifest,
real-check-*-m2d and tools/extract-codex-versions.mjs.
